]> git.karo-electronics.de Git - mv-sheeva.git/blobdiff - lib/xz/xz_dec_lzma2.c
Merge tag 'v2.6.38' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[mv-sheeva.git] / lib / xz / xz_dec_lzma2.c
diff --git a/lib/xz/xz_dec_lzma2.c b/lib/xz/xz_dec_lzma2.c
new file mode 100644 (file)
index 0000000..ea5fa4f
--- /dev/null
@@ -0,0 +1,1171 @@
+/*
+ * LZMA2 decoder
+ *
+ * Authors: Lasse Collin <lasse.collin@tukaani.org>
+ *          Igor Pavlov <http://7-zip.org/>
+ *
+ * This file has been put into the public domain.
+ * You can do whatever you want with this file.
+ */
+
+#include "xz_private.h"
+#include "xz_lzma2.h"
+
+/*
+ * Range decoder initialization eats the first five bytes of each LZMA chunk.
+ */
+#define RC_INIT_BYTES 5
+
+/*
+ * Minimum number of usable input buffer to safely decode one LZMA symbol.
+ * The worst case is that we decode 22 bits using probabilities and 26
+ * direct bits. This may decode at maximum of 20 bytes of input. However,
+ * lzma_main() does an extra normalization before returning, thus we
+ * need to put 21 here.
+ */
+#define LZMA_IN_REQUIRED 21
+
+/*
+ * Dictionary (history buffer)
+ *
+ * These are always true:
+ *    start <= pos <= full <= end
+ *    pos <= limit <= end
+ *
+ * In multi-call mode, also these are true:
+ *    end == size
+ *    size <= size_max
+ *    allocated <= size
+ *
+ * Most of these variables are size_t to support single-call mode,
+ * in which the dictionary variables address the actual output
+ * buffer directly.
+ */
+struct dictionary {
+       /* Beginning of the history buffer */
+       uint8_t *buf;
+
+       /* Old position in buf (before decoding more data) */
+       size_t start;
+
+       /* Position in buf */
+       size_t pos;
+
+       /*
+        * How full dictionary is. This is used to detect corrupt input that
+        * would read beyond the beginning of the uncompressed stream.
+        */
+       size_t full;
+
+       /* Write limit; we don't write to buf[limit] or later bytes. */
+       size_t limit;
+
+       /*
+        * End of the dictionary buffer. In multi-call mode, this is
+        * the same as the dictionary size. In single-call mode, this
+        * indicates the size of the output buffer.
+        */
+       size_t end;
+
+       /*
+        * Size of the dictionary as specified in Block Header. This is used
+        * together with "full" to detect corrupt input that would make us
+        * read beyond the beginning of the uncompressed stream.
+        */
+       uint32_t size;
+
+       /*
+        * Maximum allowed dictionary size in multi-call mode.
+        * This is ignored in single-call mode.
+        */
+       uint32_t size_max;
+
+       /*
+        * Amount of memory currently allocated for the dictionary.
+        * This is used only with XZ_DYNALLOC. (With XZ_PREALLOC,
+        * size_max is always the same as the allocated size.)
+        */
+       uint32_t allocated;
+
+       /* Operation mode */
+       enum xz_mode mode;
+};
+
+/* Range decoder */
+struct rc_dec {
+       uint32_t range;
+       uint32_t code;
+
+       /*
+        * Number of initializing bytes remaining to be read
+        * by rc_read_init().
+        */
+       uint32_t init_bytes_left;
+
+       /*
+        * Buffer from which we read our input. It can be either
+        * temp.buf or the caller-provided input buffer.
+        */
+       const uint8_t *in;
+       size_t in_pos;
+       size_t in_limit;
+};
+
+/* Probabilities for a length decoder. */
+struct lzma_len_dec {
+       /* Probability of match length being at least 10 */
+       uint16_t choice;
+
+       /* Probability of match length being at least 18 */
+       uint16_t choice2;
+
+       /* Probabilities for match lengths 2-9 */
+       uint16_t low[POS_STATES_MAX][LEN_LOW_SYMBOLS];
+
+       /* Probabilities for match lengths 10-17 */
+       uint16_t mid[POS_STATES_MAX][LEN_MID_SYMBOLS];
+
+       /* Probabilities for match lengths 18-273 */
+       uint16_t high[LEN_HIGH_SYMBOLS];
+};
+
+struct lzma_dec {
+       /* Distances of latest four matches */
+       uint32_t rep0;
+       uint32_t rep1;
+       uint32_t rep2;
+       uint32_t rep3;
+
+       /* Types of the most recently seen LZMA symbols */
+       enum lzma_state state;
+
+       /*
+        * Length of a match. This is updated so that dict_repeat can
+        * be called again to finish repeating the whole match.
+        */
+       uint32_t len;
+
+       /*
+        * LZMA properties or related bit masks (number of literal
+        * context bits, a mask dervied from the number of literal
+        * position bits, and a mask dervied from the number
+        * position bits)
+        */
+       uint32_t lc;
+       uint32_t literal_pos_mask; /* (1 << lp) - 1 */
+       uint32_t pos_mask;         /* (1 << pb) - 1 */
+
+       /* If 1, it's a match. Otherwise it's a single 8-bit literal. */
+       uint16_t is_match[STATES][POS_STATES_MAX];
+
+       /* If 1, it's a repeated match. The distance is one of rep0 .. rep3. */
+       uint16_t is_rep[STATES];
+
+       /*
+        * If 0, distance of a repeated match is rep0.
+        * Otherwise check is_rep1.
+        */
+       uint16_t is_rep0[STATES];
+
+       /*
+        * If 0, distance of a repeated match is rep1.
+        * Otherwise check is_rep2.
+        */
+       uint16_t is_rep1[STATES];
+
+       /* If 0, distance of a repeated match is rep2. Otherwise it is rep3. */
+       uint16_t is_rep2[STATES];
+
+       /*
+        * If 1, the repeated match has length of one byte. Otherwise
+        * the length is decoded from rep_len_decoder.
+        */
+       uint16_t is_rep0_long[STATES][POS_STATES_MAX];
+
+       /*
+        * Probability tree for the highest two bits of the match
+        * distance. There is a separate probability tree for match
+        * lengths of 2 (i.e. MATCH_LEN_MIN), 3, 4, and [5, 273].
+        */
+       uint16_t dist_slot[DIST_STATES][DIST_SLOTS];
+
+       /*
+        * Probility trees for additional bits for match distance
+        * when the distance is in the range [4, 127].
+        */
+       uint16_t dist_special[FULL_DISTANCES - DIST_MODEL_END];
+
+       /*
+        * Probability tree for the lowest four bits of a match
+        * distance that is equal to or greater than 128.
+        */
+       uint16_t dist_align[ALIGN_SIZE];
+
+       /* Length of a normal match */
+       struct lzma_len_dec match_len_dec;
+
+       /* Length of a repeated match */
+       struct lzma_len_dec rep_len_dec;
+
+       /* Probabilities of literals */
+       uint16_t literal[LITERAL_CODERS_MAX][LITERAL_CODER_SIZE];
+};
+
+struct lzma2_dec {
+       /* Position in xz_dec_lzma2_run(). */
+       enum lzma2_seq {
+               SEQ_CONTROL,
+               SEQ_UNCOMPRESSED_1,
+               SEQ_UNCOMPRESSED_2,
+               SEQ_COMPRESSED_0,
+               SEQ_COMPRESSED_1,
+               SEQ_PROPERTIES,
+               SEQ_LZMA_PREPARE,
+               SEQ_LZMA_RUN,
+               SEQ_COPY
+       } sequence;
+
+       /* Next position after decoding the compressed size of the chunk. */
+       enum lzma2_seq next_sequence;
+
+       /* Uncompressed size of LZMA chunk (2 MiB at maximum) */
+       uint32_t uncompressed;
+
+       /*
+        * Compressed size of LZMA chunk or compressed/uncompressed
+        * size of uncompressed chunk (64 KiB at maximum)
+        */
+       uint32_t compressed;
+
+       /*
+        * True if dictionary reset is needed. This is false before
+        * the first chunk (LZMA or uncompressed).
+        */
+       bool need_dict_reset;
+
+       /*
+        * True if new LZMA properties are needed. This is false
+        * before the first LZMA chunk.
+        */
+       bool need_props;
+};
+
+struct xz_dec_lzma2 {
+       /*
+        * The order below is important on x86 to reduce code size and
+        * it shouldn't hurt on other platforms. Everything up to and
+        * including lzma.pos_mask are in the first 128 bytes on x86-32,
+        * which allows using smaller instructions to access those
+        * variables. On x86-64, fewer variables fit into the first 128
+        * bytes, but this is still the best order without sacrificing
+        * the readability by splitting the structures.
+        */
+       struct rc_dec rc;
+       struct dictionary dict;
+       struct lzma2_dec lzma2;
+       struct lzma_dec lzma;
+
+       /*
+        * Temporary buffer which holds small number of input bytes between
+        * decoder calls. See lzma2_lzma() for details.
+        */
+       struct {
+               uint32_t size;
+               uint8_t buf[3 * LZMA_IN_REQUIRED];
+       } temp;
+};
+
+/**************
+ * Dictionary *
+ **************/
+
+/*
+ * Reset the dictionary state. When in single-call mode, set up the beginning
+ * of the dictionary to point to the actual output buffer.
+ */
+static void dict_reset(struct dictionary *dict, struct xz_buf *b)
+{
+       if (DEC_IS_SINGLE(dict->mode)) {
+               dict->buf = b->out + b->out_pos;
+               dict->end = b->out_size - b->out_pos;
+       }
+
+       dict->start = 0;
+       dict->pos = 0;
+       dict->limit = 0;
+       dict->full = 0;
+}
+
+/* Set dictionary write limit */
+static void dict_limit(struct dictionary *dict, size_t out_max)
+{
+       if (dict->end - dict->pos <= out_max)
+               dict->limit = dict->end;
+       else
+               dict->limit = dict->pos + out_max;
+}
+
+/* Return true if at least one byte can be written into the dictionary. */
+static inline bool dict_has_space(const struct dictionary *dict)
+{
+       return dict->pos < dict->limit;
+}
+
+/*
+ * Get a byte from the dictionary at the given distance. The distance is
+ * assumed to valid, or as a special case, zero when the dictionary is
+ * still empty. This special case is needed for single-call decoding to
+ * avoid writing a '\0' to the end of the destination buffer.
+ */
+static inline uint32_t dict_get(const struct dictionary *dict, uint32_t dist)
+{
+       size_t offset = dict->pos - dist - 1;
+
+       if (dist >= dict->pos)
+               offset += dict->end;
+
+       return dict->full > 0 ? dict->buf[offset] : 0;
+}
+
+/*
+ * Put one byte into the dictionary. It is assumed that there is space for it.
+ */
+static inline void dict_put(struct dictionary *dict, uint8_t byte)
+{
+       dict->buf[dict->pos++] = byte;
+
+       if (dict->full < dict->pos)
+               dict->full = dict->pos;
+}
+
+/*
+ * Repeat given number of bytes from the given distance. If the distance is
+ * invalid, false is returned. On success, true is returned and *len is
+ * updated to indicate how many bytes were left to be repeated.
+ */
+static bool dict_repeat(struct dictionary *dict, uint32_t *len, uint32_t dist)
+{
+       size_t back;
+       uint32_t left;
+
+       if (dist >= dict->full || dist >= dict->size)
+               return false;
+
+       left = min_t(size_t, dict->limit - dict->pos, *len);
+       *len -= left;
+
+       back = dict->pos - dist - 1;
+       if (dist >= dict->pos)
+               back += dict->end;
+
+       do {
+               dict->buf[dict->pos++] = dict->buf[back++];
+               if (back == dict->end)
+                       back = 0;
+       } while (--left > 0);
+
+       if (dict->full < dict->pos)
+               dict->full = dict->pos;
+
+       return true;
+}
+
+/* Copy uncompressed data as is from input to dictionary and output buffers. */
+static void dict_uncompressed(struct dictionary *dict, struct xz_buf *b,
+                             uint32_t *left)
+{
+       size_t copy_size;
+
+       while (*left > 0 && b->in_pos < b->in_size
+                       && b->out_pos < b->out_size) {
+               copy_size = min(b->in_size - b->in_pos,
+                               b->out_size - b->out_pos);
+               if (copy_size > dict->end - dict->pos)
+                       copy_size = dict->end - dict->pos;
+               if (copy_size > *left)
+                       copy_size = *left;
+
+               *left -= copy_size;
+
+               memcpy(dict->buf + dict->pos, b->in + b->in_pos, copy_size);
+               dict->pos += copy_size;
+
+               if (dict->full < dict->pos)
+                       dict->full = dict->pos;
+
+               if (DEC_IS_MULTI(dict->mode)) {
+                       if (dict->pos == dict->end)
+                               dict->pos = 0;
+
+                       memcpy(b->out + b->out_pos, b->in + b->in_pos,
+                                       copy_size);
+               }
+
+               dict->start = dict->pos;
+
+               b->out_pos += copy_size;
+               b->in_pos += copy_size;
+       }
+}
+
+/*
+ * Flush pending data from dictionary to b->out. It is assumed that there is
+ * enough space in b->out. This is guaranteed because caller uses dict_limit()
+ * before decoding data into the dictionary.
+ */
+static uint32_t dict_flush(struct dictionary *dict, struct xz_buf *b)
+{
+       size_t copy_size = dict->pos - dict->start;
+
+       if (DEC_IS_MULTI(dict->mode)) {
+               if (dict->pos == dict->end)
+                       dict->pos = 0;
+
+               memcpy(b->out + b->out_pos, dict->buf + dict->start,
+                               copy_size);
+       }
+
+       dict->start = dict->pos;
+       b->out_pos += copy_size;
+       return copy_size;
+}
+
+/*****************
+ * Range decoder *
+ *****************/
+
+/* Reset the range decoder. */
+static void rc_reset(struct rc_dec *rc)
+{
+       rc->range = (uint32_t)-1;
+       rc->code = 0;
+       rc->init_bytes_left = RC_INIT_BYTES;
+}
+
+/*
+ * Read the first five initial bytes into rc->code if they haven't been
+ * read already. (Yes, the first byte gets completely ignored.)
+ */
+static bool rc_read_init(struct rc_dec *rc, struct xz_buf *b)
+{
+       while (rc->init_bytes_left > 0) {
+               if (b->in_pos == b->in_size)
+                       return false;
+
+               rc->code = (rc->code << 8) + b->in[b->in_pos++];
+               --rc->init_bytes_left;
+       }
+
+       return true;
+}
+
+/* Return true if there may not be enough input for the next decoding loop. */
+static inline bool rc_limit_exceeded(const struct rc_dec *rc)
+{
+       return rc->in_pos > rc->in_limit;
+}
+
+/*
+ * Return true if it is possible (from point of view of range decoder) that
+ * we have reached the end of the LZMA chunk.
+ */
+static inline bool rc_is_finished(const struct rc_dec *rc)
+{
+       return rc->code == 0;
+}
+
+/* Read the next input byte if needed. */
+static __always_inline void rc_normalize(struct rc_dec *rc)
+{
+       if (rc->range < RC_TOP_VALUE) {
+               rc->range <<= RC_SHIFT_BITS;
+               rc->code = (rc->code << RC_SHIFT_BITS) + rc->in[rc->in_pos++];
+       }
+}
+
+/*
+ * Decode one bit. In some versions, this function has been splitted in three
+ * functions so that the compiler is supposed to be able to more easily avoid
+ * an extra branch. In this particular version of the LZMA decoder, this
+ * doesn't seem to be a good idea (tested with GCC 3.3.6, 3.4.6, and 4.3.3
+ * on x86). Using a non-splitted version results in nicer looking code too.
+ *
+ * NOTE: This must return an int. Do not make it return a bool or the speed
+ * of the code generated by GCC 3.x decreases 10-15 %. (GCC 4.3 doesn't care,
+ * and it generates 10-20 % faster code than GCC 3.x from this file anyway.)
+ */
+static __always_inline int rc_bit(struct rc_dec *rc, uint16_t *prob)
+{
+       uint32_t bound;
+       int bit;
+
+       rc_normalize(rc);
+       bound = (rc->range >> RC_BIT_MODEL_TOTAL_BITS) * *prob;
+       if (rc->code < bound) {
+               rc->range = bound;
+               *prob += (RC_BIT_MODEL_TOTAL - *prob) >> RC_MOVE_BITS;
+               bit = 0;
+       } else {
+               rc->range -= bound;
+               rc->code -= bound;
+               *prob -= *prob >> RC_MOVE_BITS;
+               bit = 1;
+       }
+
+       return bit;
+}
+
+/* Decode a bittree starting from the most significant bit. */
+static __always_inline uint32_t rc_bittree(struct rc_dec *rc,
+                                          uint16_t *probs, uint32_t limit)
+{
+       uint32_t symbol = 1;
+
+       do {
+               if (rc_bit(rc, &probs[symbol]))
+                       symbol = (symbol << 1) + 1;
+               else
+                       symbol <<= 1;
+       } while (symbol < limit);
+
+       return symbol;
+}
+
+/* Decode a bittree starting from the least significant bit. */
+static __always_inline void rc_bittree_reverse(struct rc_dec *rc,
+                                              uint16_t *probs,
+                                              uint32_t *dest, uint32_t limit)
+{
+       uint32_t symbol = 1;
+       uint32_t i = 0;
+
+       do {
+               if (rc_bit(rc, &probs[symbol])) {
+                       symbol = (symbol << 1) + 1;
+                       *dest += 1 << i;
+               } else {
+                       symbol <<= 1;
+               }
+       } while (++i < limit);
+}
+
+/* Decode direct bits (fixed fifty-fifty probability) */
+static inline void rc_direct(struct rc_dec *rc, uint32_t *dest, uint32_t limit)
+{
+       uint32_t mask;
+
+       do {
+               rc_normalize(rc);
+               rc->range >>= 1;
+               rc->code -= rc->range;
+               mask = (uint32_t)0 - (rc->code >> 31);
+               rc->code += rc->range & mask;
+               *dest = (*dest << 1) + (mask + 1);
+       } while (--limit > 0);
+}
+
+/********
+ * LZMA *
+ ********/
+
+/* Get pointer to literal coder probability array. */
+static uint16_t *lzma_literal_probs(struct xz_dec_lzma2 *s)
+{
+       uint32_t prev_byte = dict_get(&s->dict, 0);
+       uint32_t low = prev_byte >> (8 - s->lzma.lc);
+       uint32_t high = (s->dict.pos & s->lzma.literal_pos_mask) << s->lzma.lc;
+       return s->lzma.literal[low + high];
+}
+
+/* Decode a literal (one 8-bit byte) */
+static void lzma_literal(struct xz_dec_lzma2 *s)
+{
+       uint16_t *probs;
+       uint32_t symbol;
+       uint32_t match_byte;
+       uint32_t match_bit;
+       uint32_t offset;
+       uint32_t i;
+
+       probs = lzma_literal_probs(s);
+
+       if (lzma_state_is_literal(s->lzma.state)) {
+               symbol = rc_bittree(&s->rc, probs, 0x100);
+       } else {
+               symbol = 1;
+               match_byte = dict_get(&s->dict, s->lzma.rep0) << 1;
+               offset = 0x100;
+
+               do {
+                       match_bit = match_byte & offset;
+                       match_byte <<= 1;
+                       i = offset + match_bit + symbol;
+
+                       if (rc_bit(&s->rc, &probs[i])) {
+                               symbol = (symbol << 1) + 1;
+                               offset &= match_bit;
+                       } else {
+                               symbol <<= 1;
+                               offset &= ~match_bit;
+                       }
+               } while (symbol < 0x100);
+       }
+
+       dict_put(&s->dict, (uint8_t)symbol);
+       lzma_state_literal(&s->lzma.state);
+}
+
+/* Decode the length of the match into s->lzma.len. */
+static void lzma_len(struct xz_dec_lzma2 *s, struct lzma_len_dec *l,
+                    uint32_t pos_state)
+{
+       uint16_t *probs;
+       uint32_t limit;
+
+       if (!rc_bit(&s->rc, &l->choice)) {
+               probs = l->low[pos_state];
+               limit = LEN_LOW_SYMBOLS;
+               s->lzma.len = MATCH_LEN_MIN;
+       } else {
+               if (!rc_bit(&s->rc, &l->choice2)) {
+                       probs = l->mid[pos_state];
+                       limit = LEN_MID_SYMBOLS;
+                       s->lzma.len = MATCH_LEN_MIN + LEN_LOW_SYMBOLS;
+               } else {
+                       probs = l->high;
+                       limit = LEN_HIGH_SYMBOLS;
+                       s->lzma.len = MATCH_LEN_MIN + LEN_LOW_SYMBOLS
+                                       + LEN_MID_SYMBOLS;
+               }
+       }
+
+       s->lzma.len += rc_bittree(&s->rc, probs, limit) - limit;
+}
+
+/* Decode a match. The distance will be stored in s->lzma.rep0. */
+static void lzma_match(struct xz_dec_lzma2 *s, uint32_t pos_state)
+{
+       uint16_t *probs;
+       uint32_t dist_slot;
+       uint32_t limit;
+
+       lzma_state_match(&s->lzma.state);
+
+       s->lzma.rep3 = s->lzma.rep2;
+       s->lzma.rep2 = s->lzma.rep1;
+       s->lzma.rep1 = s->lzma.rep0;
+
+       lzma_len(s, &s->lzma.match_len_dec, pos_state);
+
+       probs = s->lzma.dist_slot[lzma_get_dist_state(s->lzma.len)];
+       dist_slot = rc_bittree(&s->rc, probs, DIST_SLOTS) - DIST_SLOTS;
+
+       if (dist_slot < DIST_MODEL_START) {
+               s->lzma.rep0 = dist_slot;
+       } else {
+               limit = (dist_slot >> 1) - 1;
+               s->lzma.rep0 = 2 + (dist_slot & 1);
+
+               if (dist_slot < DIST_MODEL_END) {
+                       s->lzma.rep0 <<= limit;
+                       probs = s->lzma.dist_special + s->lzma.rep0
+                                       - dist_slot - 1;
+                       rc_bittree_reverse(&s->rc, probs,
+                                       &s->lzma.rep0, limit);
+               } else {
+                       rc_direct(&s->rc, &s->lzma.rep0, limit - ALIGN_BITS);
+                       s->lzma.rep0 <<= ALIGN_BITS;
+                       rc_bittree_reverse(&s->rc, s->lzma.dist_align,
+                                       &s->lzma.rep0, ALIGN_BITS);
+               }
+       }
+}
+
+/*
+ * Decode a repeated match. The distance is one of the four most recently
+ * seen matches. The distance will be stored in s->lzma.rep0.
+ */
+static void lzma_rep_match(struct xz_dec_lzma2 *s, uint32_t pos_state)
+{
+       uint32_t tmp;
+
+       if (!rc_bit(&s->rc, &s->lzma.is_rep0[s->lzma.state])) {
+               if (!rc_bit(&s->rc, &s->lzma.is_rep0_long[
+                               s->lzma.state][pos_state])) {
+                       lzma_state_short_rep(&s->lzma.state);
+                       s->lzma.len = 1;
+                       return;
+               }
+       } else {
+               if (!rc_bit(&s->rc, &s->lzma.is_rep1[s->lzma.state])) {
+                       tmp = s->lzma.rep1;
+               } else {
+                       if (!rc_bit(&s->rc, &s->lzma.is_rep2[s->lzma.state])) {
+                               tmp = s->lzma.rep2;
+                       } else {
+                               tmp = s->lzma.rep3;
+                               s->lzma.rep3 = s->lzma.rep2;
+                       }
+
+                       s->lzma.rep2 = s->lzma.rep1;
+               }
+
+               s->lzma.rep1 = s->lzma.rep0;
+               s->lzma.rep0 = tmp;
+       }
+
+       lzma_state_long_rep(&s->lzma.state);
+       lzma_len(s, &s->lzma.rep_len_dec, pos_state);
+}
+
+/* LZMA decoder core */
+static bool lzma_main(struct xz_dec_lzma2 *s)
+{
+       uint32_t pos_state;
+
+       /*
+        * If the dictionary was reached during the previous call, try to
+        * finish the possibly pending repeat in the dictionary.
+        */
+       if (dict_has_space(&s->dict) && s->lzma.len > 0)
+               dict_repeat(&s->dict, &s->lzma.len, s->lzma.rep0);
+
+       /*
+        * Decode more LZMA symbols. One iteration may consume up to
+        * LZMA_IN_REQUIRED - 1 bytes.
+        */
+       while (dict_has_space(&s->dict) && !rc_limit_exceeded(&s->rc)) {
+               pos_state = s->dict.pos & s->lzma.pos_mask;
+
+               if (!rc_bit(&s->rc, &s->lzma.is_match[
+                               s->lzma.state][pos_state])) {
+                       lzma_literal(s);
+               } else {
+                       if (rc_bit(&s->rc, &s->lzma.is_rep[s->lzma.state]))
+                               lzma_rep_match(s, pos_state);
+                       else
+                               lzma_match(s, pos_state);
+
+                       if (!dict_repeat(&s->dict, &s->lzma.len, s->lzma.rep0))
+                               return false;
+               }
+       }
+
+       /*
+        * Having the range decoder always normalized when we are outside
+        * this function makes it easier to correctly handle end of the chunk.
+        */
+       rc_normalize(&s->rc);
+
+       return true;
+}
+
+/*
+ * Reset the LZMA decoder and range decoder state. Dictionary is nore reset
+ * here, because LZMA state may be reset without resetting the dictionary.
+ */
+static void lzma_reset(struct xz_dec_lzma2 *s)
+{
+       uint16_t *probs;
+       size_t i;
+
+       s->lzma.state = STATE_LIT_LIT;
+       s->lzma.rep0 = 0;
+       s->lzma.rep1 = 0;
+       s->lzma.rep2 = 0;
+       s->lzma.rep3 = 0;
+
+       /*
+        * All probabilities are initialized to the same value. This hack
+        * makes the code smaller by avoiding a separate loop for each
+        * probability array.
+        *
+        * This could be optimized so that only that part of literal
+        * probabilities that are actually required. In the common case
+        * we would write 12 KiB less.
+        */
+       probs = s->lzma.is_match[0];
+       for (i = 0; i < PROBS_TOTAL; ++i)
+               probs[i] = RC_BIT_MODEL_TOTAL / 2;
+
+       rc_reset(&s->rc);
+}
+
+/*
+ * Decode and validate LZMA properties (lc/lp/pb) and calculate the bit masks
+ * from the decoded lp and pb values. On success, the LZMA decoder state is
+ * reset and true is returned.
+ */
+static bool lzma_props(struct xz_dec_lzma2 *s, uint8_t props)
+{
+       if (props > (4 * 5 + 4) * 9 + 8)
+               return false;
+
+       s->lzma.pos_mask = 0;
+       while (props >= 9 * 5) {
+               props -= 9 * 5;
+               ++s->lzma.pos_mask;
+       }
+
+       s->lzma.pos_mask = (1 << s->lzma.pos_mask) - 1;
+
+       s->lzma.literal_pos_mask = 0;
+       while (props >= 9) {
+               props -= 9;
+               ++s->lzma.literal_pos_mask;
+       }
+
+       s->lzma.lc = props;
+
+       if (s->lzma.lc + s->lzma.literal_pos_mask > 4)
+               return false;
+
+       s->lzma.literal_pos_mask = (1 << s->lzma.literal_pos_mask) - 1;
+
+       lzma_reset(s);
+
+       return true;
+}
+
+/*********
+ * LZMA2 *
+ *********/
+
+/*
+ * The LZMA decoder assumes that if the input limit (s->rc.in_limit) hasn't
+ * been exceeded, it is safe to read up to LZMA_IN_REQUIRED bytes. This
+ * wrapper function takes care of making the LZMA decoder's assumption safe.
+ *
+ * As long as there is plenty of input left to be decoded in the current LZMA
+ * chunk, we decode directly from the caller-supplied input buffer until
+ * there's LZMA_IN_REQUIRED bytes left. Those remaining bytes are copied into
+ * s->temp.buf, which (hopefully) gets filled on the next call to this
+ * function. We decode a few bytes from the temporary buffer so that we can
+ * continue decoding from the caller-supplied input buffer again.
+ */
+static bool lzma2_lzma(struct xz_dec_lzma2 *s, struct xz_buf *b)
+{
+       size_t in_avail;
+       uint32_t tmp;
+
+       in_avail = b->in_size - b->in_pos;
+       if (s->temp.size > 0 || s->lzma2.compressed == 0) {
+               tmp = 2 * LZMA_IN_REQUIRED - s->temp.size;
+               if (tmp > s->lzma2.compressed - s->temp.size)
+                       tmp = s->lzma2.compressed - s->temp.size;
+               if (tmp > in_avail)
+                       tmp = in_avail;
+
+               memcpy(s->temp.buf + s->temp.size, b->in + b->in_pos, tmp);
+
+               if (s->temp.size + tmp == s->lzma2.compressed) {
+                       memzero(s->temp.buf + s->temp.size + tmp,
+                                       sizeof(s->temp.buf)
+                                               - s->temp.size - tmp);
+                       s->rc.in_limit = s->temp.size + tmp;
+               } else if (s->temp.size + tmp < LZMA_IN_REQUIRED) {
+                       s->temp.size += tmp;
+                       b->in_pos += tmp;
+                       return true;
+               } else {
+                       s->rc.in_limit = s->temp.size + tmp - LZMA_IN_REQUIRED;
+               }
+
+               s->rc.in = s->temp.buf;
+               s->rc.in_pos = 0;
+
+               if (!lzma_main(s) || s->rc.in_pos > s->temp.size + tmp)
+                       return false;
+
+               s->lzma2.compressed -= s->rc.in_pos;
+
+               if (s->rc.in_pos < s->temp.size) {
+                       s->temp.size -= s->rc.in_pos;
+                       memmove(s->temp.buf, s->temp.buf + s->rc.in_pos,
+                                       s->temp.size);
+                       return true;
+               }
+
+               b->in_pos += s->rc.in_pos - s->temp.size;
+               s->temp.size = 0;
+       }
+
+       in_avail = b->in_size - b->in_pos;
+       if (in_avail >= LZMA_IN_REQUIRED) {
+               s->rc.in = b->in;
+               s->rc.in_pos = b->in_pos;
+
+               if (in_avail >= s->lzma2.compressed + LZMA_IN_REQUIRED)
+                       s->rc.in_limit = b->in_pos + s->lzma2.compressed;
+               else
+                       s->rc.in_limit = b->in_size - LZMA_IN_REQUIRED;
+
+               if (!lzma_main(s))
+                       return false;
+
+               in_avail = s->rc.in_pos - b->in_pos;
+               if (in_avail > s->lzma2.compressed)
+                       return false;
+
+               s->lzma2.compressed -= in_avail;
+               b->in_pos = s->rc.in_pos;
+       }
+
+       in_avail = b->in_size - b->in_pos;
+       if (in_avail < LZMA_IN_REQUIRED) {
+               if (in_avail > s->lzma2.compressed)
+                       in_avail = s->lzma2.compressed;
+
+               memcpy(s->temp.buf, b->in + b->in_pos, in_avail);
+               s->temp.size = in_avail;
+               b->in_pos += in_avail;
+       }
+
+       return true;
+}
+
+/*
+ * Take care of the LZMA2 control layer, and forward the job of actual LZMA
+ * decoding or copying of uncompressed chunks to other functions.
+ */
+XZ_EXTERN enum xz_ret xz_dec_lzma2_run(struct xz_dec_lzma2 *s,
+                                      struct xz_buf *b)
+{
+       uint32_t tmp;
+
+       while (b->in_pos < b->in_size || s->lzma2.sequence == SEQ_LZMA_RUN) {
+               switch (s->lzma2.sequence) {
+               case SEQ_CONTROL:
+                       /*
+                        * LZMA2 control byte
+                        *
+                        * Exact values:
+                        *   0x00   End marker
+                        *   0x01   Dictionary reset followed by
+                        *          an uncompressed chunk
+                        *   0x02   Uncompressed chunk (no dictionary reset)
+                        *
+                        * Highest three bits (s->control & 0xE0):
+                        *   0xE0   Dictionary reset, new properties and state
+                        *          reset, followed by LZMA compressed chunk
+                        *   0xC0   New properties and state reset, followed
+                        *          by LZMA compressed chunk (no dictionary
+                        *          reset)
+                        *   0xA0   State reset using old properties,
+                        *          followed by LZMA compressed chunk (no
+                        *          dictionary reset)
+                        *   0x80   LZMA chunk (no dictionary or state reset)
+                        *
+                        * For LZMA compressed chunks, the lowest five bits
+                        * (s->control & 1F) are the highest bits of the
+                        * uncompressed size (bits 16-20).
+                        *
+                        * A new LZMA2 stream must begin with a dictionary
+                        * reset. The first LZMA chunk must set new
+                        * properties and reset the LZMA state.
+                        *
+                        * Values that don't match anything described above
+                        * are invalid and we return XZ_DATA_ERROR.
+                        */
+                       tmp = b->in[b->in_pos++];
+
+                       if (tmp >= 0xE0 || tmp == 0x01) {
+                               s->lzma2.need_props = true;
+                               s->lzma2.need_dict_reset = false;
+                               dict_reset(&s->dict, b);
+                       } else if (s->lzma2.need_dict_reset) {
+                               return XZ_DATA_ERROR;
+                       }
+
+                       if (tmp >= 0x80) {
+                               s->lzma2.uncompressed = (tmp & 0x1F) << 16;
+                               s->lzma2.sequence = SEQ_UNCOMPRESSED_1;
+
+                               if (tmp >= 0xC0) {
+                                       /*
+                                        * When there are new properties,
+                                        * state reset is done at
+                                        * SEQ_PROPERTIES.
+                                        */
+                                       s->lzma2.need_props = false;
+                                       s->lzma2.next_sequence
+                                                       = SEQ_PROPERTIES;
+
+                               } else if (s->lzma2.need_props) {
+                                       return XZ_DATA_ERROR;
+
+                               } else {
+                                       s->lzma2.next_sequence
+                                                       = SEQ_LZMA_PREPARE;
+                                       if (tmp >= 0xA0)
+                                               lzma_reset(s);
+                               }
+                       } else {
+                               if (tmp == 0x00)
+                                       return XZ_STREAM_END;
+
+                               if (tmp > 0x02)
+                                       return XZ_DATA_ERROR;
+
+                               s->lzma2.sequence = SEQ_COMPRESSED_0;
+                               s->lzma2.next_sequence = SEQ_COPY;
+                       }
+
+                       break;
+
+               case SEQ_UNCOMPRESSED_1:
+                       s->lzma2.uncompressed
+                                       += (uint32_t)b->in[b->in_pos++] << 8;
+                       s->lzma2.sequence = SEQ_UNCOMPRESSED_2;
+                       break;
+
+               case SEQ_UNCOMPRESSED_2:
+                       s->lzma2.uncompressed
+                                       += (uint32_t)b->in[b->in_pos++] + 1;
+                       s->lzma2.sequence = SEQ_COMPRESSED_0;
+                       break;
+
+               case SEQ_COMPRESSED_0:
+                       s->lzma2.compressed
+                                       = (uint32_t)b->in[b->in_pos++] << 8;
+                       s->lzma2.sequence = SEQ_COMPRESSED_1;
+                       break;
+
+               case SEQ_COMPRESSED_1:
+                       s->lzma2.compressed
+                                       += (uint32_t)b->in[b->in_pos++] + 1;
+                       s->lzma2.sequence = s->lzma2.next_sequence;
+                       break;
+
+               case SEQ_PROPERTIES:
+                       if (!lzma_props(s, b->in[b->in_pos++]))
+                               return XZ_DATA_ERROR;
+
+                       s->lzma2.sequence = SEQ_LZMA_PREPARE;
+
+               case SEQ_LZMA_PREPARE:
+                       if (s->lzma2.compressed < RC_INIT_BYTES)
+                               return XZ_DATA_ERROR;
+
+                       if (!rc_read_init(&s->rc, b))
+                               return XZ_OK;
+
+                       s->lzma2.compressed -= RC_INIT_BYTES;
+                       s->lzma2.sequence = SEQ_LZMA_RUN;
+
+               case SEQ_LZMA_RUN:
+                       /*
+                        * Set dictionary limit to indicate how much we want
+                        * to be encoded at maximum. Decode new data into the
+                        * dictionary. Flush the new data from dictionary to
+                        * b->out. Check if we finished decoding this chunk.
+                        * In case the dictionary got full but we didn't fill
+                        * the output buffer yet, we may run this loop
+                        * multiple times without changing s->lzma2.sequence.
+                        */
+                       dict_limit(&s->dict, min_t(size_t,
+                                       b->out_size - b->out_pos,
+                                       s->lzma2.uncompressed));
+                       if (!lzma2_lzma(s, b))
+                               return XZ_DATA_ERROR;
+
+                       s->lzma2.uncompressed -= dict_flush(&s->dict, b);
+
+                       if (s->lzma2.uncompressed == 0) {
+                               if (s->lzma2.compressed > 0 || s->lzma.len > 0
+                                               || !rc_is_finished(&s->rc))
+                                       return XZ_DATA_ERROR;
+
+                               rc_reset(&s->rc);
+                               s->lzma2.sequence = SEQ_CONTROL;
+
+                       } else if (b->out_pos == b->out_size
+                                       || (b->in_pos == b->in_size
+                                               && s->temp.size
+                                               < s->lzma2.compressed)) {
+                               return XZ_OK;
+                       }
+
+                       break;
+
+               case SEQ_COPY:
+                       dict_uncompressed(&s->dict, b, &s->lzma2.compressed);
+                       if (s->lzma2.compressed > 0)
+                               return XZ_OK;
+
+                       s->lzma2.sequence = SEQ_CONTROL;
+                       break;
+               }
+       }
+
+       return XZ_OK;
+}
+
+XZ_EXTERN struct xz_dec_lzma2 *xz_dec_lzma2_create(enum xz_mode mode,
+                                                  uint32_t dict_max)
+{
+       struct xz_dec_lzma2 *s = kmalloc(sizeof(*s), GFP_KERNEL);
+       if (s == NULL)
+               return NULL;
+
+       s->dict.mode = mode;
+       s->dict.size_max = dict_max;
+
+       if (DEC_IS_PREALLOC(mode)) {
+               s->dict.buf = vmalloc(dict_max);
+               if (s->dict.buf == NULL) {
+                       kfree(s);
+                       return NULL;
+               }
+       } else if (DEC_IS_DYNALLOC(mode)) {
+               s->dict.buf = NULL;
+               s->dict.allocated = 0;
+       }
+
+       return s;
+}
+
+XZ_EXTERN enum xz_ret xz_dec_lzma2_reset(struct xz_dec_lzma2 *s, uint8_t props)
+{
+       /* This limits dictionary size to 3 GiB to keep parsing simpler. */
+       if (props > 39)
+               return XZ_OPTIONS_ERROR;
+
+       s->dict.size = 2 + (props & 1);
+       s->dict.size <<= (props >> 1) + 11;
+
+       if (DEC_IS_MULTI(s->dict.mode)) {
+               if (s->dict.size > s->dict.size_max)
+                       return XZ_MEMLIMIT_ERROR;
+
+               s->dict.end = s->dict.size;
+
+               if (DEC_IS_DYNALLOC(s->dict.mode)) {
+                       if (s->dict.allocated < s->dict.size) {
+                               vfree(s->dict.buf);
+                               s->dict.buf = vmalloc(s->dict.size);
+                               if (s->dict.buf == NULL) {
+                                       s->dict.allocated = 0;
+                                       return XZ_MEM_ERROR;
+                               }
+                       }
+               }
+       }
+
+       s->lzma.len = 0;
+
+       s->lzma2.sequence = SEQ_CONTROL;
+       s->lzma2.need_dict_reset = true;
+
+       s->temp.size = 0;
+
+       return XZ_OK;
+}
+
+XZ_EXTERN void xz_dec_lzma2_end(struct xz_dec_lzma2 *s)
+{
+       if (DEC_IS_MULTI(s->dict.mode))
+               vfree(s->dict.buf);
+
+       kfree(s);
+}